Lysimeter based evaporation and condensation dynamics in a Mediterranean ecosystem

The input of liquid water to terrestrial ecosystems is composed of rain and non-rainfall water input (NRWI). The latter comprises dew, fog, and adsorption of atmospheric vapor on soil particle surfaces. Although NRWIs can be relevant to support ecosystem functioning in seasonally dry ecosystems, they are understudied, being relatively small, and therefore hard to measure. In this study, we test a routine for analyzing lysimeter data specifically to determine NRWI. We apply it on one year of 5 data from large high-precision weighing lysimeters at a semi-arid Mediterranean site and quantify that NRWIs occur for at least 3 h on 297 days (81 % of the year) with a mean diel duration of 6 hours.They reflect a pronounced seasonality as modulated by environmental conditions (i.e., temperature and net radiation). During the wet season, both dew and fog dominate NRWI, while during the dry season it is soil adsorption of atmospheric vapor. Although NRWI contributes only 7.4 % to the annual water input NRWI is the only water input to the ecosystem during 15 weeks, mainly in the dry season. Benefitting from the 10 comprehensive set of measurements at the Majadas instrumental site, we show that our findings are in line with (i) independent model simulations forced with (near-) surface energy and moisture measurements and (ii) eddy covariance-derived latent heat flux estimates. This study shows that NRWI can be reliably quantified through high-resolution weighing lysimeters and a few additional measurements. Their main occurrence during night-time underlines the necessity to consider ecosystem water fluxes at high temporal resolution and with 24-hour coverage. 15


Introduction
Water availability at the land surface controls a variety of processes related to land-atmosphere exchange, such as the warming of the surface and air temperature (T a , • C) (Seneviratne et al., 2010;Panwar et al., 2019), ecosystem carbon fluxes (Reichstein et al., 2007;El-Madany et al., 2021), and evapotranspiration (ET , mm) (Jung et al., 2010;Rodriguez-Iturbe et al., temperature and moisture content between the inside of the lysimeter and the surrounding soil to prevent biases (Groh et al., 2016). Therefore, data collected with large weighing lysimeters can further contribute to the identification and quantification of NRWIs. Yet, relatively few stations are located in semi-arid and arid environments Dijkema et al., 2018;Kohfahl et al., 2019;Zhang et al., 2019b) where NRWIs are expected to be particularly relevant.
In this study, we implement and test a processing scheme for identifying and quantifying NRWIs in a seasonally dry ecosys-70 tem in continental Spain. The aims of this paper are to (i) extend and refine processing routines for water flux partitioning to distinguish between ET, rain, and individual NRWIs, based on lysimeter and meteorological data; (ii) analyze the seasonal NRWI dynamics and their contribution to the annual water input at the site; and (iii) evaluate our results against independent observations and empirical models.

Study site
All data investigated in this study originates from the experimental field site Majadas de Tietar, Caceres in Extremadura, Spain (39°56 25.12 N, 5°46 28.70 W). Average diel T a is 16.7 • C with diel minimum and maximum T a of 3.1 • C to 12.5 • C in January and 18.6 • C to 39.8 • C in August. The rain mainly falls between October and April, with mean annual amounts of ca. 650 mm, with large interannual variations (El-Madany et al., 2020). The ecosystem is a typical Mediterranean semi-80 arid tree-grass ecosystem (Dehesa) with low-density oak tree cover (Quercus Ilex (L.), 20 trees ha −1 ) (Bogdanovich et al., 2021). The herbaceous layer consists of native annual grasses, forbs, and legumes  with a seasonally varying fractional cover mainly influenced by moisture availability (Luo et al., 2020). The growing season for the herbaceous layer begins after the first rains after summer (typically in mid-October) and is inhibited by low temperatures in winter before peaking in spring before the dry season. During the dry season, the herbaceous species are inactive until the return of rain 85 . The site is managed with low-intensity grazing by cows during the growing season (El-Madany et al., 2018). An exclusion cage was used to avoid cows stepping into the lysimeters. However, the structure of the cage allowed for grazing to maintain the lysimeters comparable with the rest of the plot. The soil is formed of alluvial deposits and classified as Abruptic Luvisol (IUSS Working Group WRB, 2015) with sandy topsoil of 74 % sand, 20 % silt, and 6 % clay (Nair et al., 2019). A clay layer rests at a variable depth between 30 cm to 100 cm. Although the trees also play a role in the water balance 90 at the ecosystem scale, herbaceous vegetation dominates ET . This work focuses on the water fluxes in open areas where lysimeters are located .

Lysimeter technical specifications
The site is equipped with three lysimeter stations, each containing two weighable high-precision lysimeters (UGT, Müncheberg, Germany), for a total of 6 columns. Each column has a 1 m 2 surface area and 1.20 m column depth. They rest of the weighing 95 systems consisting of three precision shear-stress cells, respectively (Model 3510, Stainless Steel Shear Beam Load Cell, VPG Transducers, Heilbronn, Germany). The weight measurements are collected every 1 min with a measurement precision of 10 g (0.01 mm). Each lysimeter station is equipped with a lower boundary control system to avoid deviations from natural conditions due to the isolation of the lysimeter columns (Groh et al., 2016). Porous ceramic bars at the bottom of the lysimeters maintain the soil water potential within the column comparable with the soil surrounding them. Soil temperature (T soil , • C) is controlled 100 with a heat exchange system (for further details, see Perez-Priego et al. (2017)). SW C and T soil are measured every 15 min within the columns at 0.1, 0.3, 0.75, and 1 m depth (UMP-1, Umwelt-Geräte-Technik GmbH, Müncheberg, Germany). Soil matric potential (Ψ, hPa) is measured every 15 min at 0.1 m depth with a porous ceramic cone full range pF meter (ecoTech Umwelt-Meßsysteme GmbH, Bonn, Germany).
The stations were installed in 2015 at open grassland patches with 104 m, 91 m, and 24 m distance to each other, respectively.

105
The closest tree is ∼9 m away. To test the application of data processing and flux classification across seasons, we analyze a period of one year from June 1st, 2019 to May 31st, 2020.

Ancillary measurements outside the lysimeters
For partitioning the lysimeter weight changes (∆W , kg min −1 ) into different water fluxes and modeling (see Section 2.2.1, 2.3.1, and 2.3.2) we used additional field measurements collected every 30 min. Meteorological variables monitored are rain, 110 which is measured with a weighing rain gauge (TRwS 514 precipitation sensor, MPS systém, Slovakia), T a , and rH at 1 m (Pt-100 capacitive humidity sensor CPK1-5, MELA Sensortechnik, Germany). T dew was calculated based on T a and rH (see Appendix A2). Actual vapor pressure (e a , hPa) is calculated from T a and rH.
T soil , and SW C were measured along a profile outside the lysimeters at 0.05 m, 0.10 m, and 0.2 m depth, respectively (Delta-ML3, Delta-T Devices Ltd, Burwell Cambridge, UK).
Fluxes of latent heat (λE, W m −2 ), wind speed (u, m s −1 ) and friction velocity (u * , m s −1 ) were measured by an eddy covariance (EC) system, consisting of a sonic anemometer (R3-50 Gill Instruments, Lymingon UK) and an infra-red gas analyzer (LI-7200, Licor Biosciences, Lincoln, USA) at 1.6 m sampling height and targeting the herbaceous layer . For further details on the EC data processing, see El-Madany et al. (2018).

Data analysis
2.2.1 Lysimeter data processing 125 The processing of lysimeter data comprises several steps: (a) raw weight data filtering, (b) time-series smoothing, and (c) flux partitioning. The processing workflow is displayed in figure 1. All code used in the analysis is available for reproducibility in the open-source R environment for statistical programming (R Core Team, 2020). See the data and code availability statement for more details.
a) Raw data processing: changes in the water reservoir through the lower boundary system and lysimeter column weights 130 are added together. Outliers are filtered out by setting plausible threshold values; −0.5 kg min −1 < ∆W < 1 kg min −1 (Schrader et al., 2013). Additionally, outliers within these threshold values were identified by comparing ∆W across the six columns. If ∆W is due to rain, we expect similar responses across lysimeters. In contrast, if only one lysimeter column shows an anomalous ∆W we considered this as an artifact (e.g., animal stepping on the column or issues with the boundary control) that can be removed from the time series (Hannes et al., 2015). For identifying these values, we calculated the mean 135 ∆W of all six lysimeters for an interval i of one minute. This value was then subtracted from the individual ∆W measurements during i. The resulting value is a normalized weight change (∆W normalized,i , kg) for each column. Then, an average standard deviation (σ, kg) was calculated from ∆W normalized,i−3 to ∆W normalized,i+3 . ∆W normalized,i > (1.5 × |σ|) are replaced by not a number (NA). b) Time-series smoothing: is necessary to remove noise from the time series before the partitioning and data analysis based 140 on ∆W (Schrader et al., 2013). Since noise in this type of data is not constant in time due to wind for example (Nolz et al., 2013), we apply a routine with adaptive averaging window widths (ω) and adaptive ∆W thresholds (δ) (AWAT) from Peters et al. (2014Peters et al. ( , 2016Peters et al. ( , 2017. As an intermediate result, the AWAT algorithm produces a step function of lysimeter weight. At last, a smoothing of the 1 min resolution time series is performed using a spline interpolation. We chose this routine because the authors included a processing step developed specifically for dew conditions (Peters et al., 2017). In our application, the 145 parameter ω varied between 3 min to 31 min and δ between 0.01 mm to 0.05 mm. A detailed overview of the algorithm and an evaluation of the performance are compiled in Peters et al. (2017) and Hannes et al. (2015). If one out of all columns was measuring more than 16 h of water input during one day, the full day was excluded for the analysis for the respective column.
Due to technical problems that became obvious after data processing, lysimeter column number 4 was completely excluded from further analysis. After the smoothing, the time series are aggregated to a 5-minute resolution to further decrease the 150 remaining influence of noise for the subsequent step of flux partitioning, particularly for values close to zero. Peters et al. 2014Peters et al. , 2016Peters et al. , 2017  c) Flux partitioning: the filtered time series of ∆W is divided into six different water fluxes, assuming that in 5-minute intervals, only one process prevails over the others. Negative ∆W is always classified as ET . Positive ∆W is separated into different flux categories in a decision tree structure considering additional meteorological data, as illustrated in figure 1. We use the convention that outgoing fluxes like ET take negative values (related to negative ∆W as water is leaving the soil), and 155 rain and NRWIs are positive (because associated with positive ∆W ). At the second decision node, we check if the rain gauge identifies a rain event during the period of weight increase. If true, all positive ∆W are classified as rain during 30 min before and after the event. This period was selected to match the measurement interval of rain gauges which is 30 minutes. If false, e.g., in the absence of rain, rH is evaluated. If rH exceeded a threshold identified as the 90 th percentile of the rH sensor records, ∆W is attributed to fog. In our case, this threshold value is at rH t = 97.8 % and measured by a sensor at 1 m height. 160 We decided to set a rH threshold that is based on the data distribution of the sensor because it accounts for the individual uncertainty which is particularly high when the air is nearly saturated (Feigenwinter et al., 2020), systematic biases, and drifts of rH sensors.
If neither fog nor rain is detected, we compared T s with T dew at 1 m sensor height (equations in Appendix A2). Note that this is not necessarily representative of T dew within the grass canopy, which would be more suitable to infer dew formation.

165
For example, through radiative cooling, the air cools faster closer to the surface than at 1 m height. To include the effect of the height difference, we compared sensors installed at 0.1 m and at 1 m height during a measurement campaign of two months in Spring 2021. The results show that the median temperature difference between the 0.1 m to 1 m sensor height is 1.
If no dew was detected, ∆W could be potentially attributed to soil adsorption of atmospheric vapor (Zhang et al., 2019a).

170
Adsorption occurs, however, at specific soil hydraulic and meteorological conditions given by a relation between Ψ and atmospheric rH. Those have been observed in the laboratory (Camuffo, 1984;Arthur et al., 2016) and suggested from field observations (Kosmas et al., 2001;Uclés et al., 2015;Zhang et al., 2019b). For adsorption Ψ falls below (or is less negative than) a given threshold. This threshold can also be expressed in terms of SW C, given the relationship between Ψ and SW C through the soil water retention curve. In order to integrate this knowledge into the classification of adsorption, measurements We test the plausibility of inferred water fluxes by three different methods: 1) qualitative and quantitative comparison against direct measurements of rain and ET and 2) comparison against model predictions of dew and adsorption in absence of respective direct measurements derived with alternative measured variables. Similar approaches were used successfully for the benchmarking of methods to simulate transpiration (Nelson et al., 2018) or carbon fluxes (Jung et al., 2020). Measured flux 195 durations were compared to respective model estimates modeled flux using correlation, mean absolute error (MAE), and root mean squared error (RMSE) (full equations Appendix A2). 3) for periods classified as fog by the partitioning algorithm we cross check with images collected by a digital camera installed at the site (Luo et al., 2018). This was, however, only applicable for few events, since the first image per day was taken at 10 am.

Modelling dew
Dew is modeled as negative latent heat flux calculated based on models originally developed for determining evapotranspiration: (i) the Penman-Monteith (PM) (Monteith, 1965) equation which combines processes related to radiative energy and vapor pressure deficit (previously applied for dew in various forms by e.g. Jacobs et al., 2006;Aguirre-Gutiérrez et al., 2019;Groh et al., 2018), and (ii) equilibrium evaporation (previously applied for dew by Uclés et al., 2014).
We implemented the models as described in Ritter et al. (2019)  .
is the density of air, and δq is the deficit 210 of specific humidity at reference level (kg kg −1 ). r av is the aerodynamic resistance to vapor transport between the surface and the air (s m −1 ) and was derived with an empirical relationship based on u * (Thom, 1972).
For both equations, dew occurs when λE < 0 and T s ≤ (T dew −1.4 • C) (as explained in section 2.2.1). This approach has been reported to be suitable for detecting potential dew conditions and to analyze dew frequency and duration. But it is limited in reproducing dew yields (Ritter et al., 2019). Hence we focus on comparing condition lengths rather than yields since we aim 215 to validate dew detection by the partitioning routine.

Modelling adsorption
Adsorption conditions were identified based on the vertical humidity gradient near the surface. We implemented the re-arranged aerodynamic diffusion equation originally used by Milly (1984) and previously applied for modeling adsorption by Verhoef et al. (2006): 220 e s,0 = γ r av λE ρ a C p + e a (2. 3) The target value e s,0 (kPa) is vapor pressure of soil air at the surface. e a (kPa) is vapor pressure of the atmosphere and C p (J kg −1 K −1 ) is the specific heat of air at constant pressure. The other parameters are the same as in Eq. (2.1). When e s,0 < e a , gradient driven vapor flow is towards the soil surface. This is assumed to be indicative of the adsorption of vapor from the atmosphere. Different from the dew models, we used high-quality filtered measurements from EC for λE in Eq. (2.3). Again,

225
we only compared the simulated daily duration of suitable conditions of atmospheric adsorption against the results obtained with the lysimeter weights partitioning and we constrained the comparison to moments when T s ≤ (T dew −1.4 • C).  shown for lysimeter column 6) and rH, respectively, together with mean diel SW C at 0.05 m depth and maximum diel T s range.
Lysimeter ∆W are mainly negative between sunrise and sunset (Fig. 2a), e.g., water is lost from the column due to ET .
After sunset, however, they are zero or positive during most of the year, indicating water input. This diel pattern is consistent across seasons, following the seasonal daylight variability. The flux classification reveals seasonal differences in the prevailing 235 NRWI (Fig. 2b). At the beginning of June, atmospheric adsorption mainly occurs during the early morning, before sunrise.
From July to September the length of the adsorption period increases, and the onset shifts towards earlier in the night. In this 9 https://doi.org/10.5194/hess-2021-519 Preprint. Discussion started: 2 November 2021 c Author(s) 2021. CC BY 4.0 License.
period, the diel variability of rH is relatively low (Fig. 2c), SW C at 5 cm depth is below 10 % and T s oscillates up to an amplitude of 35 • C day −1 (Fig. 2d & e). A rain event, in late July increases SW C and is followed by some days of increased ET which also prevails during night-time.

240
A rain event in late September leads to longer-lasting increases in SW C and rH. Such conditions are typically associated with vegetation re-greening. ET and dew alternate during night-time. Frequent rain events in November and December are accompanied by dew and fog becoming the dominant NRWIs until the end of the measurement period in April.
NRWI occur for at least 3 h during 297 days (81 % of the year) with a mean duration of 6 hours per day of occurrence.
The weekly sums in Fig. 3 illustrate that the seasonal dynamics are consistent across lysimeters. The ecosystem receives 245 atmospheric water at any time of the year but with shifting relative relevance of the water flux types over the year. Rain is the dominant liquid water input (i.e. it contributes more than 50 % of the weekly water input) during 29 weeks since it's total amount is usually much greater than NRWI, whereas NRWI is dominant in 24 weeks. They are even the only water input during 15 weeks with adsorption as exclusive water input in 10 weeks of the year. Dew and fog occurrence is synchronized with rain with regard to the seasonal occurrence, and therefore their relative contribution is small. The median contribution of adsorption 250 until September 2019 is 0.9 mm week −1 . With ET amounting to −5.7 mm week −1 , thus adsorption compensates for 19 % of the weekly water loss during summer.  and lysimeter and EC (bottom). The points and vertical error bars indicate mean and standard deviation between lysimeter columns, which include measurement uncertainty and spatial variability.

Water flux sums and their consistency with measurements and theory
The cumulative measured rain is 565.0 ± 11 mm as an average across the six lysimeters. Thereby, the difference between the maximum and minimum across the lysimeters is 30.7 mm, which is an absolute deviation of 5 % between columns. For 255 comparison, the rain gauge recorded 597 mm of rain during the same time period. The underestimation of the estimates from lysimeters compared to the rain gauge, as well as the deviation between columns, is mainly caused by a few large rain events in January 2020 (Fig. 4a). The cumulative measured ET across lysimeters is −570.7 ± 20 mm. The annual cumulated difference between lysimeters is 46.7 mm. Annual cumulative ET determined through EC measurements is −619 mm. As in the case of rain, the ET estimates deviate strongest in winter which indicates a technical problem during days with rain at this time of the 260 year, while estimates are more consistent during the rest of the study period (Fig. 4d).
Descriptive statistics of the annual sums of NRWIs across lysimeter columns are summarised in adsorption. The differences in flux sums across lysimeters for all NRWI are relatively larger than for rain and ET , with coefficients of variation around 40 %. The largest relative annual deviation between columns was found for adsorption with 265 a difference of 20.0 mm. But the flux which is most affected by the threshold parameter in the lysimeter flux partitioning is fog, which can be seen in Table 2. The IQR for the fluxes decreases the later in the partitioning scheme the respective flux is estimated (Fig. 1). Thereby, for dew and adsorption, the spatial variability between lysimeter columns (IQR in Table 1) exceeds the range of uncertainty related to the partitioning parameters (IQR in Table 2).
To assess whether the thermodynamic requirements for dew and adsorption are met at our site we compare the lysimeter-270 inferred observations of dew and adsorption with their potential occurrence determined with the models from PM, equilibrium evaporation, and the aerodynamic diffusion equation (Fig. 5). Our results show that the measured fluxes are temporally consistent with model results, both concerning diurnal and seasonal dynamics.
In general, the Majadas site has suitable conditions for dew between October 2019 and end of May 2020, from sunset to sunrise. The statistical metrics for the comparison of daily dew duration between lysimeters and models are summarized in 275   Table A2. They show that overall, the models suggest a longer duration of dew conditions by 3 h day −1 to 5 h day −1 . Model statistics from PM and the evaporation model are not deviating from each other indicating that in our application, no difference between the simplified and full PM model is detectable. In the case of adsorption the lysimeter-based estimates agree better with the model predictions (Table A2). When comparing only measurements where at least two out of the five lysimeters show weight increases assigned to adsorption, evaluation statistics improve by one hour. In addition, the agreement of lysimeters is 280 overall stronger during adsorption than during dew conditions (Fig. 5). Single lysimeters, however, frequently also measured adsorption until midday and before sunset.

Discussion
In this study, we showed that large weighing lysimeters and a few ancillary measurements can be used to efficiently disentangle and therefore quantify all types of surface water fluxes, particularly NRWI. Our data from a semi-arid Mediterranean savanna 285 site shows that the climatic conditions fulfill the thermodynamic requirements to induce diel cycles of evaporation and condensation at almost all times of the year. The routine proposed to detect and distinguish NRWI is successfully validated against models based on energy balance and moisture gradients, regarding the occurrence of the process. We found that NRWIs occur frequently at our site, in line with previous research in such a climate regime. The occurrence of both adsorption and dew was shown by Zhang et al. (2019b). We support their observation that dew formation and ad-290 sorption dominate at different times of the year. Regular dew formation (120 nights year −1 to 200 nights year −1 ) has been reported across sites (Tomaszkiewicz et al., 2015). In a similar semi-arid steppe ecosystem in Spain, the mean number of days per year with suitable conditions for dew formation was 285 days (Uclés et al., 2014), however, their ecosystem is close to the Mediterranean sea and likely, therefore, receives more humidity than Majadas. Our observation that especially nights are prone to the formation of NRWI is also documented in the literature. Dew formation length has often been reported to 295 correlate with the length of the night (Tomaszkiewicz et al., 2015) and was in another Spanish site reported to last on average 9.3 ± 3.2 hours night −1 (Uclés et al., 2014). In contrast, for adsorption, reported observation times differ. Kosmas et al. (1998) observed the flux to occur mainly between 0:00 and 6:00 h and also Saaltink et al. (2020) found suitable night-time conditions for adsorption through a reversed gradient of vapor concentration between soil and atmosphere from lysimeter observations and confirmed it with a fully coupled numerical model. Yet, Verhoef et al. (2006) found adsorption occurring during the afternoon 300 and ceasing at night. Since changes in T s and their effects on the phase equilibrium of water are one of the main controlling factors of adsorption, the different findings between studies could be related to site-specific timing of surface exposure to radiation.
Next to the diagnosed occurrences, also the NRWI amounts we determined are within the range of previously reported estimates for similar climate regimes. At our study site Majadas, the largest annual NRWI contribution is adsorption with  (Zhang et al., 2019b;Kidron and Starinsky, 2019). It is important to remark, however, that part of the large variability concerning the length of occurrence and condensation rates for NRWI could be related to biases of the measurement devices. Micro lysimeters, until now one of the most widely used instruments, were reported by Kidron and Kronenfeld (2020b) to overestimate NRWI likely due to greater heat loss through the walls, compared to the surrounding unperturbed soil. 315 We found differences in the absolute annual NRWI sums between individual lysimeters that can be attributed either to (i) spatial variability (heterogeneity) in the soil and vegetation characteristics affecting the energy balance, or (ii) instrumental and methodological uncertainty. Particularly for the latter, external disturbances by wind or animals and internal disturbances such as data gaps, and data processing have been shown to alter results significantly (Schrader et al., 2013;Nolz et al., 2013).
Developing processing routines for raw data from large weighing lysimeters has challenged researchers during the last decade 320 (Schrader et al., 2013;Peters et al., 2014Peters et al., , 2016Peters et al., , 2017Hannes et al., 2015). To assess the robustness of the processing we compared rain across lysimeters as rain is expected to be similar across nearby lysimeters and heterogeneous soils and vegetation. The variation of rain across lysimeters is only 5 %, which is of a similar magnitude reported in other studies (Hannes et al., 2015;Schneider et al., 2021). Model simulations independently confirmed, that the conditions at night are suitable for NRWI, although especially for dew the simulations showed that the potential for dew formation is generally longer than actual 325 occurrence, measured with lysimeters.
A focus of this study was partitioning the lysimeter ∆W into water flux classes. This approach includes the simplified assumption that one flux is always dominating over the others at each time step. In reality, the fluxes can occur simultaneously with their relative importance shifting gradually over time (Li et al., 2021b), but postulate to be minimal at the time scale we are looking at. Ideally, we would account for a statistical probability ratio between different NRWI per time interval. But current 330 research that is quantifying such ratios is too scarce for generalization (Li et al., 2021b).

Disentangling individual NRWI fluxes is nevertheless important because of different respective i) controlling factors and
ii) implications for the ecosystem. Better knowledge on controlling factors can help to identify potential NRWI occurrence also in ecosystems without specialized measurement devices. The role of dew has often been reported as moistening plant surfaces with direct leaf water uptake (Tomaszkiewicz et al., 2015). Soil vapor adsorption, however, occurs at low Ψ where 335 grasses in Majadas have already senesced. Nevertheless, due to it's continuous occurrence in periods of low SW C, it would be a beneficial ecological strategy for organisms to use this diel water input. So far, it is recognized that microbial and lichen respiration is triggered by adsorption (Evans et al., 2019;McHugh et al., 2015;Dirks et al., 2010;Li et al., 2021a;Gliksman et al., 2017) but more research is necessary to understand the impact of this flux on different organisms. The approach applied most frequently in the literature as well as in this study is based on a discontinuous tree-based classification system which was 340 implemented similarly as suggested by Zhang et al. (2019a). As an extension of their system one node was added to account for prior knowledge on the controls of adsorption. The prior knowledge for soil adsorption stems, however, from smaller samples at equilibrium conditions in the laboratory (Arthur et al., 2016). Our application to point measurements of rH and SW C from above and below the soil surface assumes that by choosing the statistical upper envelope we can distinguish equilibrium conditions from remaining noise in the time series as best as possible. Although uncertainty remains about the real shape of 345 this relationship, this approach gives a more conservative estimate of the adsorption amount and helps prevent overestimation.
The validity of this relationship is further confirmed by having a similar shape independently whether it was derived from the periods when lysimeter measurements unanimously were classified as adsorption (before including residuals as a final node) or deduced from negative EC derived λE.
The advantage of the discontinuous tree-based classification is that it is applicable widely because the necessary data is 350 commonly measured. The disadvantage is that selection of the parameters and thresholds in the classification algorithm is critical, especially at the upper nodes, where choices propagate into the estimates of flux classes of deeper nodes. Our test on different parameter combinations found that the strongest impact on flux quantity was for fog, followed by adsorption.
Nevertheless, the ranking was not affected by choice of parameters and thresholds, e.g. in all tested cases adsorption had the largest contribution to annual flux sums.

355
Apart from instrumental and methodological uncertainty, the spatial heterogeneity of soil and vegetation characteristics of the field site (Nair et al., 2019) can affect our results. In fact, dew and adsorption amounts are both reported to vary substantially with the surface cover type (Uclés et al., 2016), soil exposure (Kosmas et al., 2001), shading hours (Uclés et al., 2015), distance from trees (Verhoef et al., 2006;Qubaja et al., 2020) and soil texture, particularly clay and sand content (Kosmas et al., 2001;Orchiston, 1952;Yamanaka and Yonetani, 1999). Verhoef et al. (2006) showed in a measurement campaign with eight 360 lysimeters concentrically arranged around a single oak tree that at the most exposed spots adsorption was doubled compared to the shaded spot. At our site, there are also individual sparse trees (Bogdanovich et al., 2021) which cause small-scale differences in shading. Since some lysimeter columns are more exposed than others, part of the deviation in NRWI could be explained by spatial heterogeneity. This is supported by measurements of soil Ψ within the individual columns, which showed that one column had an overall greater mean T soil in summer and the threshold of potential adsorption conditions was reached 365 nearly a full month earlier than in other columns. Micrometeorological variables were however not measured individually at each column and therefore we have no insight into the exact causes of spatial heterogeneity in dew formation. The applied models both only suggest times of dew formation potential based on measurements at the central facility, while the quantity is derived from weight changes. For the same reason, spatial differences are robust and follow-up investigations can be targeted towards understanding their causes.

370
Notably, the current study also provides an opportunity to propose a way forward in adsorption research based on the observed similar temporal pattern between negative EC derived λE and lysimeter vapor adsorption occurrence. EC instruments are currently one of the most popular instruments to estimate λE at an ecosystem scale (Baldocchi, 2014(Baldocchi, , 2020. But measurements are frequently discarded when the underlying micrometeorological assumptions of the technology are not met (Göckede et al., 2004). This often affects night-time EC measurements (Massman and Lee, 2002). Previous research in a pine forest in 375 Israel also indicated that EC-derived λE tends to be negative at night during adsorption (Qubaja et al., 2020). At our site, this pattern is obvious and indicates that night-time EC measurements could serve to detect adsorption (Fig. 6).
To help assess the relevance of adsorption, we suggest revaluing night-time λE fluxes of EC instruments. This could also help to up scaling and in doing so overcoming the problems with clarifying the role of NRWI across research communities (Gerlein-Safdi, 2021), for example concerning energy balance closure of EC (de Roode et al., 2010), ecological significance of 380 foliar water uptake (Berry et al., 2019) and impacts on remote sensing products (Xu et al., 2021). Irrespectively, our findings underline the necessity for methods on processes of the water cycle during the night to avoid biased measurements towards evaporation while missing condensation.

Conclusions and outlook
In this manuscript, we derive NRWI from time series of automated weighted lysimeters and compare their length of occurrence 385 with established model estimations. In summary, our data suggest that this semi-arid savanna ecosystem switches between evaporation and condensation almost daily. Attributing the condensation pattern into different NRWI sheds light on the distinct mechanisms that are each dominant in a different season. In summer, adsorption of atmospheric vapor on soil particles is facilitated by large diel temperature differences and dry soils lead to steep gradients of atmospheric vapor pressure between the atmosphere and the soil air driving vapor diffusion. In winter and spring, high rH leads to frequent fog deposition and surface 390 cooling to dew condensation. Spatial heterogeneity of vegetation, soil characteristics, and radiation regimes, together with measurement uncertainties are stronger reflected in NRWI than in other fluxes of surface water exchange. Although between 01.06.2019 and 31.05.2020, the total NRWI sum comprises only 7.4 % of the local water input, the relative contribution strongly varies weekly. Rain frequency is unevenly distributed within the year and especially atmospheric adsorption stands out as the only water input during 11 weeks in the dry season. The ecological relevance of this flux has yet to be scrutinized.

395
Our analysis focuses on lysimeters, that cover only a spatial area of 1 m 2 , each. We show that the temporal variability of the NRWI derived from the instruments is coherent with negative LE fluxes at dry conditions. Based on this observation, future work could focus on revalue night-time λE fluxes from EC instruments to improve the spatial representativeness and assess the relevance of NRWI at the larger scale and across seasonally dry ecosystems.
Dewpoint temperature (T dew , • C) was calculated from rH and T a based on the Magnus equation (λ = 17.62, β = 243.12) (Sonntag, 1990): where rH is relative humidity (%) and T a is air temperature ( • C).
Evaluation statistics the comparison of the occurrence duration between lysimeter measurements and modeling